Ionospheric Stinger - euler.nmt.edueuler.nmt.edu/~brian/students/jstinger.pdf · Chapman PIM Fit...

Nonlinear Inversion of �� A Airglow to Ionospheric O�

Number Density

by

Jon F� Stinger

Submitted in Partial Ful�llment

of the Requirements for the Degree of

Master of Science in Mathematics with Operations Research Option

New Mexico Institute of Mining and Technology

Socorro� New Mexico

December ��

ACKNOWLEDGEMENT

I would like to thank Dr� Borchers� without whose constant attention

this project would probably never have been �nished� I would like to thank

Dr� Thomas for all his help and support� and I would like to thank both Dr�

Thomas and his associates at the NRL for the opportunity to be a part of this

project� I would like to thank the rest of my committee� Dr� Hossain and Dr�

Schaer� for their help and support� I would like to thank the rest of the Math

Department at New Mexico Tech for everything over the last couple of years�

Finally� I would like to thank my family� friends� roommates� and others who

have put up with me over the time that this project has occupied my existence�

for being supporting and patient through the whole thing�

This thesis was typeset with LATEX� by the author�

�LATEX document preparation system was developed by Leslie Lamport as a special version

of Donald Knuth�s TEX program for computer typesetting� TEX is a trademark of the

American Mathematical Society� The LATEX macro package for the New Mexico Institute of

Mining and Technology thesis format was adapted by Gerald Arnold from the LATEX macro

package for The University of Texas at Austin by Khe�Sing The�

ii

ABSTRACT

Inverse methods are often required to solve geophysical problems in

which we estimate real geophysical conditions from measurements taken from

a remote sensor� Often� we can estimate what measurements would result from

a given set of conditions using physical equations� or forward equations� Nearly

as often� it is very dicult to solve the inverse of these physical equations� In

these cases we are required to estimate the parameters of these forward equa�

tions by estimating and changing these parameters until we �nd a reasonable

�t� To �nd a reasonable set of parameters� we use an inverse method such as

the Levenberg�Marquardt method of least�squares parameter estimation� In

this thesis� we use the Levenberg�Marquardt method to estimate O� number

densities in the ionosphere given readings of �� A ultraviolet radiation taken

from satellite instruments�

Table of Contents

Acknowledgement ii

Abstract

Table of Contents iii

List of Tables v

List of Figures vi

�� Introduction �

�� Nonlinear Inverse Methods �

�� Levenberg�Marquardt Method of Nonlinear Least Squares Pa�

rameter Estimation � � � � � � � � � � � � � � � � � � � � � � � � � �

�� The Steepest Descent Method � � � � � � � � � � � � � � � �

�� The Gauss�Newton Method � � � � � � � � � � � � � � � � �

�� The Levenberg�Marquardt Method � � � � � � � � � � � � ��

�� Tikhonov Regularization � � � � � � � � � � � � � � � � � � � � � � ��

�� The L�Curve � � � � � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Inverse Model of O� Number Density Using the Chapman

Equation ��

�� Forward Model Using the Chapman Equation � � � � � � � � � � ��

iii

�� Chapman Inverse Model Using an Initial Ionospheric Model Es�

timated from the Chapman Equation � � � � � � � � � � � � � � � ��

�� Chapman Inverse Model Using an Initial Ionospheric Model Es�

timated from PIM � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Inverse Model of O� Number Density Using a Spline Fit Model �

�� The Spline Fit � � � � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Forward Model Using the Spline Fit Model � � � � � � � � � � � � ��

�� Spline Inverse Model Using an Initial Ionospheric Model Esti�

mated from the Chapman Equation � � � � � � � � � � � � � � � � ��

�� Spline Inverse Model Using an Initial Ionospheric Model Esti�

mated from PIM � � � � � � � � � � � � � � � � � � � � � � � � � � ��

� Results and Conclusions �

References ��

iv

List of Tables

�� Multiplication Factors of Parameters in Chapman Equation � � ��

�� Constants Used in Forward Model Equation � � � � � � � � � � � ��

�� Constants Used in the MSIS�� Process Used in the Forward

Model � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Constants Used in Forward Model Equation � � � � � � � � � � � ��

�� Chapman to Chapman Models Using Varying Noise Final Pa�

rameters � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Chapman to Chapman Models Using Varying Noise Tests � � � ��

�� Chapman to PIM Models Using Varying Noise Final Parameters ��

�� Chapman to PIM Models Using Varying Noise Tests � � � � � � ��

�� Chapman to PIM Models Using Varying Noise Final Parameters

With A Quarter the Noise � � � � � � � � � � � � � � � � � � � � � ��

�� Chapman to PIM Models Using Varying Noise Tests With A

Quarter the Noise � � � � � � � � � � � � � � � � � � � � � � � � � � ��

v

List of Figures

�� Limbscan Diagram � � � � � � � � � � � � � � � � � � � � � � � � � �

�� Example L�curve � � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Chapman Fit of Chapman Data � � � � � � � � � � � � � � � � � � ��

�� Chapman Forward Model Fit to Chapman Data � � � � � � � � � ��

�� Chapman�Chapman Standardized Residuals � � � � � � � � � � � ��

�� Chapman�Chapman Normal Probability Plot � � � � � � � � � � � ��

�� Chapman�Chapman Regression of Residuals � � � � � � � � � � � ��

�� Chapman�Chapman Fit Noise � � � � � � � � � � � � � � � � � � � ��









�� Chapman�Chapman Residuals Noise � � � � � � � � � � � � � � � ��

vi









�� Chapman Fit of PIM Data � � � � � � � � � � � � � � � � � � � � � ��

�� Chapman Forward Model Fit to PIM Data � � � � � � � � � � � � ��

�� Chapman�PIM Standardized Residuals � � � � � � � � � � � � � � ��

�� Chapman�PIM Normal Probability Plot � � � � � � � � � � � � � ��

�� Chapman�PIM Regression of Residuals � � � � � � � � � � � � � � ��

�� Chapman�PIM Fit Noise � � � � � � � � � � � � � � � � � � � � � � ��








vii


�� Chapman�PIM Residuals Noise � � � � � � � � � � � � � � � � � � ��









�� Chapman�PIM Fit Quarter Noise � � � � � � � � � � � � � � � � � ��









�� Chapman�PIM Residuals Quarter Noise � � � � � � � � � � � � � ��


viii








�� Chapman L Curve � � � � � � � � � � � � � � � � � � � � � � � � � ��

�� Spline Chapman Fit ��





�� Spline Chapman Standardized Residuals � � � � � � � � � � � � � ��

�� Spline�PIM Fit L Curve � � � � � � � � � � � � � � � � � � � � � � ��

�� Spline�PIM Fit ��





ix






x

Chapter �

Introduction

Geophysical problems often require the use of inverse methods to �nd

solutions� An inverse problem is a problem in which we know the outcome

of readings taken from a sensor or set of sensors of some kind� What we are

hoping to �nd� however� is the geophysical state which produces the readings

obtained by the sensor�

We can physically develop an equation that will allow us to generate what read�

ings will be produced from certain geophysical states� This is what is called a

forward problem� A forward problem progresses directly forward in a physically

logical way from initial physical conditions to results� That is� given a set of

geophysical conditions� we can derive an equation that predicts the resulting

readings from the sensor� The procedure for solving an inverse problem uses

the procedure for solving the forward problem to determine the initial condi�

tions�

We can often use optimization methods such as the Levenberg�Marquardt

Method �Levenberg� �� Marquardt� �� Lawson and Hanson� �� Bard� ��

Gill et al�� Bevington and Robinson� �� with methods like Tikhonov

Regularization �Tikhonov and Arsenin� �� to �nd the model that best es�

timates the retrieved data with respect to a least squared error �t� The

best possible model would be geophysically feasible� and as simple as possi�

�

�

ble �Parker� �� We �nd this best �t by minimizing �� the sum of the

squared errors between the set of retrieved data� d� and the data produced

from our estimated parameters� �d� divided by the variance of each parameter�

�� as follows�

�� nXi��

� �di � di��

��i��

�Bevington and Robinson� ��

The geophysical problem of interest is the determination of the O� ion number

density� or atoms per unit volume� in the F layer of the ionosphere� The F

layer of the ionosphere is between the altitudes of �� and �� kilometers�

This measurement is made at a spacecraft by the reception of �� A radia�

tion in the ultraviolet portion of the electromagnetic spectrum� A number of

upcoming satellite projects will gather �� A airglow data in the interest of

modelling the O� number density in the F layer of the ionosphere� These

projects include the Remote Atmospheric And Ionospheric Detection System

�RAIDS� �Christiansen et al�� and the Special Sensor Ultraviolet Limb

Imager �SSULI� �McCoy et al��

When solar radiation strikes an oxygen atom in and below the F layer of the

ionosphere� the molecule absorbs enough energy to allow the oxygen to ion�

ize� In the ionization process� an electron is released� along with a photon at

�� A� These photons are of a wavelength to allow multiple resonant scatter�

ing� Multiple resonant scattering is the absorbtion and reemission of photons

by atoms� which changes the direction of motion for the said photon� This

means that all photons moving along a speci�c line of sight were not necessar�

ily originally emitted from O� ions along that line of sight�Picone et al��

McCoy et al�� Additionally� other ions and molecules in the ionosphere

�

can simply absorb free photons� These factors complicate the determination of

O� number density from �� A ultraviolet radiation intensity �Picone et al��

These satellites will use a technique called limb�scanning� Limb�scanning is a

process in which measurements are taken at many altitudes in the F layer� by

varying the angle at which the sensor looks� The angles of these measurements

range from skimming the very top of the ionosphere to pointing vertically into

the ionosphere� Data over multiple angles allow for estimations to be made for

multiple layers of the ionosphere� A �rst reading would �nd an intensity for

only the very top �layer� of the ionosphere� A second reading� with an angle

penetrating into more of the ionosphere� would �nd an intensity reading due

to two layers� including the �rst� Measurement at a number of angles� then�

gives us intensities due to many layers of the ionosphere� These data would be

produced from a set of speci�c O� number densities in the forward model� We

can therefore use these readings in an inverse model to �nd a good model set

of number densities that would produce given readings�

A forward model that translates O� number densities in the ionosphere to

�� A ultraviolet intensity is necessary to solve the desired inverse problem�

To solve the forward problem� we need to know the conditions in the iono�

sphere at the time the measurements were taken� These conditions include

the concentrations of other molecules and atoms �N�� O�� N� He� O� H� and

Ar�� and temperature� For the generation of this data� and data in previous

research �Picone et al�� we use the mass spectrometer incoherent scatter

�� MSIS�� model �Hedin� �� MSIS�� generates this data from infor�

mation including� year� day� latitude� longitude� universal time� solar zenith

angle� and solar ��cm �ux from the previous day and averaged over the pre�

�

��

��

��

��

��

Satellite

Earth

F layer of Ionosphere

Figure �� Diagram of Limb Scanning of the F Layer of the Ionosphere Sepa�rated into Layers

vious �� days� These values are found separately at the time of the satellite

readings and are used in the MSIS�� process as constants�

�Picone et al�� modelled the O� number density of the F layer of the

ionosphere using the Levenberg�Marquardt method for nonlinear least squares

problems in �� The satellite data would be used to �t a four parameter

equation that modelled observed ionospheric data� the four�parameter Chap�

man equation �Anderson and Meyer� �� The four parameters coincide with�

the altitude of maximum number density� the number density at that altitude�

and two parameters that control the drop�o rates of the number density as

altitude varies� These four parameters� given satellite data� are estimated us�

ing the Levenberg�Marquardt method� Unfortunately� the Chapman equation�

while it accurately models most currently available ionospheric O� number

density� does not work well in all cases and may reject interesting features� It

�

would� therefore� be bene�cial to have an inverse procedure that does not use

the Chapman equation�

There are other models that are used for estimating O� concentrations in the

ionosphere� such as the parameterized ionospheric model �PIM�� This iono�

spheric model does not use the four�parameter Chapman equation� but instead

forms data sets from theoretical equations based on parameters including mag�

netic activity� interplanetary magnetic �elds� ��cm solar �ux� season� and

longitude �Daniell Jr� et al�� Like the Chapman equation� these param�

eters establish a pro�le over the F layer of the ionosphere�

The purpose of this thesis is to establish a means of determining the O� con�

centrations over the F layer of the ionosphere� This model should use only

estimated concentrations at given altitudes with a cubic spline curve interpo�

lating between those altitudes to make a �t to data from a satellite� This model

would accomplish the goal of eschewing the Chapman equation� or any other

pro�le�generating process� while creating a reasonable� smooth model of the F

layer of the ionosphere�

While this model would be more acceptable in the �eld of ionospheric research�

it is mathematically more complicated because of its ill�conditioned nature�

An ill�conditioned problem is one in which small changes to values in the prob�

lem can cause large changes in the solution�Gill et al�� If we use the

Levenberg�Marquardt method by itself in this and other ill�conditioned prob�

lems� the resulting solution set is rough� increasing and decreasing erratically�

This erratic model is physically unreasonable� The layer parameters should be

smooth� increasing quickly and evenly then decreasing more slowly as we rise

through the given altitudes� To �nd a smooth solution to this inverse prob�

�

lem using the Levenberg�Marquardt method requires the use of regularization�

which will smooth the solution� making it physically reasonable� In this case�

Tikhonov regularization is used to smooth the solution�

The problem of uniqueness arises in inverse theory� When using parameters

to generate a model to represent the retrieved data� we can possibly get an

in�nite number of solutions� Is there a unique solution� To �nd a solution to

the problem of parameter �tting� we minimize an objective function based on

�nding a good �t to the retrieved data� We know that there is noise in the

retrieved data� so we do not wish to match it exactly� What we wish to do is

�nd a smooth function that minimizes the sum of squared errors between the

retrieved data and the �tted data� If we are regularizing� we minimize the sum

of squared errors plus the regularization term� We have uniqueness because

this minimum is the �t that best minimizes this objective function�

Chapter �

Nonlinear Inverse Methods

�� Levenberg Marquardt Method of Nonlinear Least Squares Pa

rameter Estimation

When solving a nonlinear inverse problem� it is often desirable to

minimize a function which is the sum of the squared errors between the values

of the retrieved data and the values of the data resulting from the chosen

parameters in the forward problem� The Levenberg�Marquardt method is a

least�squares method of parameter estimation for nonlinear problems� It is an

iterative method in which we use an initial estimated set of model parameters

to generate a �nal set of model parameters that minimizes the sum of squared

errors between itself and the data set retrieved from research�

These methods take advantage of the fact that residual errors for a number of

data points of a function have some interesting and useful properties� Given

residual errors between the retrieved data set� �d� and the data set derived from

initial guess parameters�d� � �di � di�� and the variance of each of those data

points �i one half of the sum of the squares of those errors is equal to the

function F �x��

F �x� ��

�

nXi��

� �di � di��

��i�

�

�k

�d� d

�k��

If f is the current set of estimated forward model solution data� the gradient

of F �x� is

rF �x� � J�x�T f�x� ��

�

�

and the Hessian is

r�F �x� � J�x�TJ�x� � Q�x� ��

where J�x� is the Jacobian of f�x�� and

Q�x� �mXi��

fi�x�r�F �x� ��

�Gill et al��

The Levenberg�Marquardt method takes advantage of two parameter estima�

tion techniques� the steepest descent method and the Gauss�Newton method�

Both of these methods have inadequacies that are resolved by the Levenberg�

Marquardt method� The steepest descent method of parameter estimation will

converge to a minimum� but does so very slowly� The Gauss�Newton method

has the opposite problem� It generally converges quickly� but sometimes fails

to converge at all� The Levenberg�Marquardt method� therefore� uses steepest

descent while the estimated parameter set is far from the best �tting parameter

set� and the Gauss�Newton method when the parameter set is very close to the

best �tting parameter set� In between� Levenberg�Marquardt uses a mixture

of both processes �Bevington and Robinson� ��

�� The Steepest Descent Method

The steepest descent method� also known as the gradient search method�

uses the negative of the gradient of the forward problem� �rF �p� at the set

of estimated parameters p� the direction of steepest descent� to �nd a direction

to search� The gradient can be estimated by �nite dierences or by using the

derivatives of the forward model with respect to the parameters if formulas for

�

the derivative are available� The negative gradient can then be multiplied by a

positive step length scalar constant � to produce a vector �p� which is added

to p to form a new set of parameters� and the process is repeated� This step

can be summarized in the equation

�p � �rF �p��

p � �p� pnew ��

�Gill et al�� Bevington and Robinson� ��

As mentioned� the problem with the steepest descent method is its slow con�

vergence rate� As we approach the set of parameters which gives a minimum

sum of squared errors when compared to the retrieved data set� the gradient

approaches zero� The change in the parameters at these values therefore also

approaches zero� At values close to the minimum� we get gradient vectors that

go one way� then another as we �ne�tune towards the minimum� At values close

to the correct minimum� we therefore get very slow convergence� The steepest

descent method converges linearly� and in practice� this linear convergence is

very slow �Gill et al�� Bevington and Robinson� ��

�� The Gauss Newton Method

Newton�s method uses the Taylor series approximation of the forward

model taken to the second derivative to �nd a step length and direction for the

parameter values p� The Taylor series expansion to the second derivative term

of the forward model is

F �x � �x� � F �x� �rF �x�T�x ��

��xTr�F �x��x� ��

��

To minimize any given function ��x�� we want its derivative r��x� � �� So

the minimum of the above expansion equation with respect to �x occurs when

its derivative with respect to �x is equal to zero�

rF �x� �r�F �x��x � ��

We can see that we get a minimum when

r�F �x��x � �rF �x� ��

�Gill et al�� Bard� �� Bevington and Robinson� ��

Examining this equation and using the above theory on least squares problems�

we can derive the Gauss�Newton method of linear least squares parameter

estimation� where �p is the step length taken on our guessed parameters that

will minimize the residuals� F �p��

�J�x�TJ�x� � Q�x��p � �J�x�T f�x��

We can in practice ignore the Q�x� in the function� as J�x�TJ�x� tends to

dominate the equation� If we remove Q�x� from the equation� we can still

get a good answer reasonably quickly under most conditions� This makes it

unnecessary to compute second derivatives� After dropping the Q�x�� we get

the Gauss�Newton method� which uses the equation

�J�x�TJ�x��p � �J�x�T f�x� ��

�Gill et al��

��

�� The Levenberg Marquardt Method

A problem with the Gauss�Newton method occurs when the J�p�TJ�p�

matrix is ill�conditioned� We need a well�conditioned J�p�TJ�p� matrix of full

rank because in the course of solving the equation

p � ��J�p�TJ�p��J�p�Tf�p��

we must factor the matrix� A poorly conditioned matrix can be impossible to

factor on a computer� as round�o can make it rank de�cient� Ill conditioning�

though� occurs all too often in nonlinear inverse problems�Gill et al��

We can make this matrix positive de�nite� and therefore Cholesky factorizable�

by adding an identity matrix multiplied by a variable element � which is large

enough to make the matrix eigenvalues strictly positive� This gives a �nal

equation of

�J�p�TJ�p� � �I��p � �J�p�Tf�p��

This equation gives us a good balance between steepest descent and Gauss�

Newton� If the J�p�TJ�p� matrix can handle a small lambda� the Levenberg�

Marquardt step is eectively a Gauss�Newton step� because �I can be negligi�

ble� If � is large enough to make the J�p�TJ�p� inconsequential in relation to

the �I� the step is eectively

��p � �J�p�T f�p� ��

a steepest descent step since it is equivalent to

�

��p � �rF �p� ��

��

where � is a scalar constant� as we have already stated that �J�p�T f�p� �

�rF �p� �Gill et al��

We therefore need a method of establishing the value of � for a given step� A

simple scheme for choosing � is as follows�

�� Let � � ��

�� Set �� to a small value� like ��

�� Compute �p and F �p � �p� using this ��

�� If F �p � �p� � F �p�� increase � by a factor of � and repeat�

�� If F �p � �p� � F �p�� decrease � by a factor of �� set p� � �p � �p� and

repeat �� using p� for p

�Marquardt� �� Bevington and Robinson� ��

By repeating this method over multiple iterations� and changing p to p � �p

only if the step is a relative improvement� we can eventually reach an optimal

solution for the parameters of the data�producing equation given the initial

data set� It is possible to reach a local minimum� so a good selection of initial

guess parameters is necessary �Gill et al��

�� Tikhonov Regularization

A regularization process is a process that is used to choose a solu�

tion that is reasonable for the intended process �Lawson and Hanson� ��

Tikhonov and Arsenin� �� Tikhonov et al�� In the case of this and

��

many other processes� reasonable can be de�ned as smooth with regard to the

second derivative of the function� As we do not get an explicit function from

an inverse model such as our spline �t model� we must utilize a regularization

procedure that uses the information we do have on the inverse model�

Tikhonov regularization is the procedure we use for the process of smooth�

ing the inverse model with respect to its second derivative� In the process of

Tikhonov regularization� we minimize the sum of two quantities� the squared

errors between the retrieved data and the produced inverse model� F �p��d� as

we would in least squares methods such as the Levenberg�Marquardt method

and the ��norm of the vector to be regularized� z� times a constant regulariza�

tion parameter � In essence� we minimize the function

kF �p�� dk� � kzk� ��

�Tikhonov and Arsenin� �� Tikhonov et al��

In the case of our problem� we wish to regularize the second derivative of the

inverse model� so we need to alter the vector to be regularized� To regularize

the second derivative of the function� we may multiply the parameters of the

inverse problem� p� by the matrix L� the matrix of sums of squares of the second

dierences� ��

��

��

��

� � ��

� � � � � � ��

��

as Lp��Twomey� �� We may then restate the equation to be minimized as

kF �p�� dk� � kLpk��

��

Figure �� Example L�Curve

This version of Tikhonov regularization compares the smoothness of the sec�

ond derivatives of the inverse function to the least squares �t of the inverse

function� The regularization parameter gives the regularization quantity of

the minimization function the weight it needs� If more smoothness is necessary�

raise the value of � If the smoothness should be less important than the value

of the least squares �t� make smaller� It can be a very important decision to

decide the value of a regularization �

�� The L Curve

The L�curve is a graphical method of choosing the best value of the

regularization parameter � The L�curve shows the tradeo between the resid�

ual norm kF �p�� dk� on the x�axis and the regularization norm kLpk� on the

y�axis� We �nd an L�curve by plotting these two values for multiple runs of

a regularization method using diering values for the regularization � These

��

points� if they include the optimal � will form an L shape� with the point

corresponding to the optimal at the corner�

At the corner� we have a that �nds the smallest value of kF �p� � dk� rela�

tive to the smallest kLpk�� The two errors are eectively balanced� If is any

larger� we get an increasingly large �nal value kLpk� while decreasing kAp�dk�

relatively little� Conversely� if is any smaller� we get an increasingly large

kAp� dk� while decreasing kLpk� relatively little �Hansen� ��

Chapter �

Inverse Model of O� Number Density Using the

Chapman Equation

My inverse model of O� number density using the Chapman equation

is based on� and is a recreation of� the model created by �Picone et al��

In many cases� this model works in the estimation of O� number density� In

some cases� though� the Chapman equation fails to accurately model the O�

number density�

We have tested the inverse model using the Chapman equation against two

ionospheric models estimated from dierent sources� one created using the

Chapman equation and one using PIM� These models are used in the absence

of real� known ionospheric O� data with corresponding �� A readings� The

Chapman equation and the PIM model estimate O� number density over the

F layer of the ionosphere� These estimated models are then run through the

forward model to produce �� A radiation data as it would be received by

satellite instruments�

We have added random noise to the data produced by the forward models to

make the estimated retrieved �� A data realistic� This random noise is based

on an error equal to a normal random variable with a standard deviation equal

to the square root of the estimated data point� This produces noise similar to

real noise in real satellite data� This is true because the satellite instruments

perform what is eectively a Poisson count of photons striking those instru�

��

��

ments� A Poisson distribution has a standard deviation � equal to the square

root of the mean � and can be acceptably estimated by a normal �or Gaussian�

distribution for a signi�cantly large �Bevington and Robinson� ��

When using the variables in the Chapman equation as parameters in our in�

verse model� no regularization is necessary� as all Chapman models are smooth

curves�

�� Forward Model Using the Chapman Equation

The four�parameter Chapman equation �Anderson and Meyer� ��

includes the following variables� the altitude of maximum number density zmax�

the number density at that altitude Nmax� and two parameters that control the

drop�o rates of the number density as altitude varies� H� and H�� These

variables are used in the four�parameter Chapman equation

NO��z� � Nmaxexp�

�

�

��

z � zmax

H� � H�z� exp

��

z � zmax

H� � H�z

��

to evaluate a number density NO� at altitudes z over the F layer of the iono�

sphere�

For purposes of the inverse procedure� we scale all of these parameters� This

scaling lowers the ill�conditioning of the Jacobian matrix by setting all of these

values to be on the order of one� In this scaling� we use parameters that are on

the order of one� and multiply them by the factors in Table �� before applying

them to the Chapman equation�

This equation is the function of a curve that estimates O� number density

over the F layer of the ionosphere� We use �fty points along this curve� at

��

Table �� Multiplication Factors of Parameters in Chapman Equation

Parameter Multiplication Factor

Nmax ��

zmax ��

H� ��

H� ��

�� km intervals over the F layer� as a discrete estimation of this curve� We

then use this discrete estimation of the O� number densities over the F layer

in our forward model to evaluate the �� A radiation readings expected from

our satellite sensors� The forward model �Picone et al�� is the following

series of equations

j�z� � j��z� � ��NO��z�Z�

z�

j�z��H�j � � j� jt� � tj�dz� ��

which evaluates the volume emission rate j�z� for �� A given the de�nitions

for the constants as given in Table �� and the MSIS�� paramenters given in

Table ��

Given this data� the volume excitation rate j�z�� we must then perform the

integration to estimate the column emission rate ��I� where I is the radiance

in megaphoton cm�� sec�� ster�� as measured at the satellite�

��I�r��e� � ��Zj�r��s��T �r�� r�ds ��

given the constants de�ned in Table ��

These integrals are discretized and evaluated numerically in code developed

for �Picone et al�� The result is an array of ninety retrieved �� A inten�

sities over varying limb�scan angles�

��

Table �� Constants Used in Forward Model Equation �

j��z� the initial photon production term ! providedby certain databases

z� altitude below which the initial photon term isnegligible or extinction is large

�� resonant scattering line�center cross sectionfor �� A radiation

�z� the vertical resonant scattering line�centeroptical depth at altitude z

t�z� the vertical pure absorbtion optical depth ataltitude z

H the Holstein probability function� which is thetransport kernal� proportional to the probabilitythat a photon propagates from one region�z�� z� � dz�� to another region �z� z � dz��

Table �� Constants Used in the MSIS�� Process Used in the Forward Model

latitude � ��

longitude � ��

local time � �� hoursday � ��year � ��

satellite altitude � �� kmsatellite inclination angle � ��

Table �� Constants Used in Forward Model Equation �

�e the line of sight of the satellite instrumentr� the initial location of the photonr the location of the satellite instrumentT the probability that a photon travelling along �e

from r� will arrive at r�

��

Figure �� Chapman�Generated O� Concentration in the F Layer of the Iono�sphere and the Chapman Equation Inverse Solution

�� Chapman Inverse Model Using an Initial Ionospheric Model

Estimated from the Chapman Equation

We �rst used the Chapman equation inverse model to �nd a solution

to a data set created by using the Chapman equation itself� The unscaled

parameters of the Chapman equation were

�zmax� Nmax� H�� H��

Our Chapman equation�based Levenberg�Marquardt program found a solution

of


These parameters produce a plot that compares to the original as shown in

Figure ��

This set of Chapman parameters produces a set of �� A satellite readings that

matches the initial set of noisy data as shown in Figure �� The Chi�squared

value of this �t is �� over �� degrees of freedom� which produces a one�sided

��

Figure �� Chapman�Generated Forward Fit �� A Intensity Data and NoisyChapman Equation Produced Data

p�value of �� Bain and Engelhardt� �� This p�value is the probability

that �� would be so large if the �t were correct� A p�value of �� means that

we have a �� probability that we would get residuals like this if this was our

initial data set� A p�value of �� or less would mean that the �t was poor�

We can also plot the standardized residuals of the forward model from our �t to

the retrieved data� This plot will show us that a good �t exists if the residuals

appear to be normally distributed� Standardized residuals are the dierence

between the retrieved data points and the �tted data divided by the estimated

standard deviation of the readings �� In this case the standard deviation is the

square root of the retrieved data� The method of determining a standardized

residual for a �t is as follows�

SRi ��di � di

�i� ��

A normally distributed set of standardized residuals would have the most val�

ues near the mean� or zero standard deviations� As we move out from the

��

Figure �� Standardized Residuals of the Chapman Equation Fit to ChapmanEquation Generated Retrieval Data

mean� the density of standardized residuals occurring would get smaller� un�

til only a few values go as far as the third standard deviation� Figure ��

shows the standardized residuals of the Chapman equation �t of the data set

generated from the Chapman equation� We also look for patterns in the resid�

uals� such as strings of sequential residuals with positive standardized residuals

�Ramsey and Schafer� ��

We may examine the standardized residuals in Figure �� and the regression

�t of these residuals in Figure �� and note that no patterns or trends appear�

We can go farther in testing the normality of the residuals by producing a nor�

mal probability plot� or Q�Q Plot� This plot is a visual aid to show normality�

If the plot is normal� the points will appear in a fairly straight diagonal line�

grouped together near the center and spread out more at the ends� If there

are curves in the plot� then the residuals are not normally distributed� We

show the normal probability plot for the �t to the Chapman data set in Figure

�� We can see from the p�value of the Anderson�Darling test that we have a

��

P-Value: 0.779A-Squared: 0.237

Anderson-Darling Normality Test

N: 90StDev: 0.950609Average: -0.0030135

210-1-2

.999

.99

.95

.80

.50

.20

.05

.01

.001

Pro

babi

lity

Chapman-Chap

Normal Probability Plot

Figure �� Anderson�Darling Normal Probability Plot for Chapman Fit toChapman Data

��

908070605040302010 0

2

1

0

-1

-2

C8

Cha

pman

-Cha

p

R-Sq = 3.8 %

Y = 0.319419 - 7.09E-03X

Regression Plot

Figure �� Regression of Residuals in Chapman to Chapman Fit

��

reasonably normal distribution of residuals �D�Augostino and Stevens� ��

Having established that we have made a reasonable �t with respect to hav�

ing normal residuals� we now must establish the likely range of the Chap�

man parameters� The parameters we have �t to the model are only esti�

mates� We need to know how certain we are with the �t� To show this cer�

tainty in the range of each parameter� we must produce a covariance matrix

�Bain and Engelhardt� �� A covariance matrix is a square� symmetric ma�

trix with a number of diagonal elements equal to the number of parameters�

Each diagonal element is the variance� �� of the respective parameter� Each

nondiagonal element is the covariance of the two elements that coincide with

its row and column� The covariance matrix established from the Levenberg�

Marquardt method is equal to �JTJ�� where J is the Jacobian of the �nal

iteration of the Levenberg�Marquardt process� The covariance matrix for the

Chapman �t of the scaled parameters of the Chapman�generated data is

��

��

��

We can use the covariance matrix to generate a correlation matrix� A correla�

tion matrix is a covariance matrix with each row i of the matrix divided by the

standard deviation of element i� �i and each column j of the matrix divided by

the standard deviation of element j� �j� This gives a matrix with ones in the

diagonal� and values between �� and � for all nondiagonal elements� Each ele�

ment shows the correlation between the parameters� A correlation of � means

that the two elements have a directly proportional relationship� If one element

increases� the other increases without changing the outcome� A correlation of

��

�� means that the two elements have an inversely proportional relationship�

One of the two elements decreases while the other increases� and the outcome

will not change �Bain and Engelhardt� �� This covariance matrix results in

the following correlation matrix

��

� ��

��

This correlation matrix shows a strong inverse correlation between the �rst two

parameters and between the last two parameters at the location of this solu�

tion� This means that one of the values in either pair can vary upward� and the

other value in that pair can vary downward� and the value of the solution would

change very little� This can cause a solution to result in a set of parameters

dierent from the original parameter set by only one of the parameters varied

up and the second in that pair varied down�

We can also use the covariance matrix to establish ��" con�dence intervals for

each parameter� A con�dence interval is an interval of values which we are ��"

con�dent that the correct answer lies within this interval �Bain and Engelhardt� ��

In the case of this solution� our ��" con�dence interval� when we have unscaled

the values from the covariance matrix� is


We provide a sample of �ts using the same initial parameters and diering

noise in Tables �� and �� and Figures �� These �ts leave standardized

residuals as shown in Figures ��

The ��" Con�dence Intervals for the parameters show the interval which is

��

Table �� Chapman to Chapman Models Using Varying Noise Final Parame�ters

Data Set zmax ��" CI Nmax ��" CI H� ��" CI H� ��" CI

� ��

��" likely to contain the correct value �Bain and Engelhardt� �� The tests

done in Table �� show how good of a �t we have� The �� and its p�value show

the likelihood of having the resulting residuals if the �t is correct� A p�value

of less than �� means that we reject the hypothesis that this is a good �t�

The p�value for the Anderson�Darling test is the likelihood of having the result�

ing residuals if the distribution of residuals is normal� A p�value of less than

�� means the residuals are not normally distributed� and the �t is not good�

The last two columns on Table �� are the ��" con�denct intervals of a linear

regression �t to the standardized residuals� The parameters should both be

zero� For a good set of residuals both con�dence intervals would contain zero�

Table �� shows that for zmax� we come close or cover the correct value of ��

with our con�dence intervals� In six of the nine trials� the correct value is not

included in the con�dence interval� All other parameters are included in all

con�dence intervals over all trials�

Table �� shows that the p�value of the �t for trial � is not good� Examining

��

Table �� Chapman to Chapman Models Using Varying Noise Tests

�� Test Regression of ResidualsData Set �� p�value Anderson� ��" CI ��" CI

Darling Regression Regressionp�value Constant x�Coe�

� ��

Figure �� we can see that the �t appears to be good� In fact� Figure �� ap�

pears to be quite possibly the best �t� But this trial is the only one which does

not pass the �� test� This shows that trusting the plot is not the appropriate

way to test goodness�of��t� We must use statistical tests to verify that the �t

is good�

Trial � on Table �� has a low Anderson�Darling p�value� This means that the

distribution of residuals is not likely to be normal� This is not terribly surpris�

ing� as we have nine trials and the p�value is likely to be less than �� even

for good �ts about �" of the time� This trial is also �awed because neither

regression coecient con�dence interval includes zero� Given this information�

we are led to believe that the residuals are not only non�normal� but their

mean is not zero� and they have a trend toward larger values as our look angle

increases�

��

Figure �� Chapman Equation Fitted to Chapman Data with Random NoiseSet �

We must also examine trials � and �� Both of these trials are acceptably normal

and have reasonable �� values� But when we examine the residual regression

�t� we see that the con�dence interval for the constant does not include zero�

The coecient of x does include zero� so there is no statistically signi�cant

trend to the residuals� But the mean of the residuals is statistically not zero�

This is not of as much concern as having a non�zero x�coecent� but must be

noted�

��



��

Figure �� Chapman Equation Fitted to Chapman Data Standardized Resid�uals with Random Noise Set �


��


�� Chapman Inverse Model Using an Initial Ionospheric Model

Estimated from PIM

We next use the Chapman parameters to model a data set generated

by a method that is not the Chapman equation� We use a set of O� num�

ber densities generated by the PIM model that are not easily modelled by the

Chapman equation� as shown in Figure �� The PIM values are the O� num�

ber densities over the altitudes of the F layer of the ionosphere�

After performing the Levenberg�Marquardt process on the initial guess param�

eters for the Chapman equation� we �nd that the unscaled Chapman equation

�t parameters are


These parameters result in the Chapman �t shown in Figure ��

The �nal Chapman parameters produce a O� pro�le that results in a forward

model �t shown in Figure �� along with the noisy data produced from the

PIM model� This model uses a standard deviation equal to the square root

��



��

Figure �� PIM�Generated O� Concentration in the F Layer of the Ionosphereand the Chapman Equation Inverse Solution

of each data point to represent instrument error� just as before� It appears

that the �t is reasonably accurate� even though the �tted Chapman pro�le is

nothing like the PIM pro�le�

To test the goodness of �t of our model of the data� we once again �nd a ��

value for the �t� The �� in this case is �� for �� degrees of freedom� This

results in a p�value of �� This p�value tells us that we cannot reject the

hypothesis that we have a good �t to the data�

We then check the standardized residuals for any patterns or deviations from

normality� Looking at Figure �� we see that the residuals appear to be nor�

mal� As another check of normality� we examine Figure �� to test normality�

The Anderson�Darling normal probability plot shows that the residuals are

generally normal� with a p�value that does not reject the hypothesis of a good

�t�

As we do not reject this model as a �t of the data� we then examine the

covariance matrix� For the Chapman �t of the PIM�generated data set� we get

��

Figure �� Chapman�Generated Forward Fit �� A Intensity Data and NoisyPIM Produced Data

Figure �� Standardized Residuals of the Chapman Equation Fit to PIMGenerated Retrieval Data

��

P-Value: 0.363A-Squared: 0.396

Anderson-Darling Normality Test

N: 90StDev: 0.980848Average: 0.133800

210-1-2

.999

.99

.95

.80

.50

.20

.05

.01

.001

Pro

babi

lity

Chapman-PIM

Normal Probability Plot

Figure �� Anderson�Darling Normal Probability Plot for Chapman Fit toPIM Data

��

908070605040302010 0

2

1

0

-1

-2

C8

Cha

pman

-PIM

R-Sq = 0.2 %

Y = 0.204335 - 1.55E-03X

Regression Plot

Figure �� Regression of Residuals in Chapman to PIM Fit

��

a covariance matrix

��

��

��

��

for the scaled Chapman parameters� This covariance matrix results in a corre�

lation matrix

��

� ��

��

These correlations show a correlation nearing �� for the last two parameters�

The last two parameters� therefore� are once again completely correlated� and

can change� one up and one down� while changing the outcome very little�

When we use the covariance matrix to establish ��" con�dence intervals for

each parameter� we get the following results�


We provide a sample of �ts using the same initial parameters and diering

noise in Tables �� and �� and Figures �� These �ts leave standardized

residuals as shown in Figures ��

From Figures �� we can see that none of these gives us a very good

estimation of the initial O� pro�le� But we have a decent �� in most cases and

pass all tests in some� We need to �nd a way to recognize a poor �t to the

correct pro�le�

We contrast the previous tables and �gures with Tables �� and �� and

��

Table �� Chapman to PIM Models Using Varying Noise Final Parameters


� ��

Figures �� Figures �� once again show residuals and any possible

trends� These tables and �gures give the results of the �tting of the same data

with random noise with a standard deviation � equal to one fourth the original

noise�

In Table �� we have little in the way of signals that we have a poor �t to

the O� pro�le� Trial � fails to pass the �� p test� Trials � and � fail the

Anderson�Darling p�test� so we can believe that the residuals of those trials are

not normally distributed� Trial � also has a constant regression of residuals

parameter that is not statistically zero� We can see from Figures �� that

none of these trials produces a good �t� But only three fail any tests at all�

These results are only slightly worse than the results of �tting to a Chapman

pro�le in Table �� The results of the tests of the �ts shown in Figures ��

�� are much better at showing that these are not good �ts� In Table �� we

see that trials �� and � all fail the �� test� Trials � and � are suspiciously

close� Going to the Anderson�Darling p�test� we can now throw out trial � for

��

Table �� Chapman to PIM Models Using Varying Noise Tests



� ��

Table �� Chapman to PIM Models Using Varying Noise Final ParametersWith A Quarter the Noise


� ��

��

Table �� Chapman to PIM Models Using Varying Noise Tests With A Quar�ter the Noise



� ��

failing to be normal� We discard �ve of nine �ts as poor �ts� as opposed to

three for full noise� A more sensitive instrument would allow us to determine

poor �ts more reliably�

��

Figure �� Chapman Equation Fitted to PIM Data with Random Noise Set�


��


Figure �� Chapman Equation Fitted to PIM Data Standardized Residualswith Random Noise Set �

��



��

Figure �� Chapman Equation Fitted to PIM Data with ��" Random NoiseSet �


��


Figure �� Chapman Equation Fitted to PIM Data Standardized Residualswith ��" Random Noise Set �

��



Chapter �

Inverse Model of O� Number Density Using a Spline Fit

Model

The goal of the spline �t inverse model is to produce an inverse proce�

dure that will model any possible ionospheric O� number density pro�le� The

spline �t model does not use the Chapman equation� and is therefore not con�

strained by its limitations� Any O� number density pro�le can theoretically be

modelled� as the parameters of the inverse procedure are points on the pro�le

itself�

We have tested the spline �t model against the same two estimated models

that we used to test the Chapman model� the Chapman equation generated

pro�le and the pro�le generated by PIM� The same noise was added to the ��

A radiation data to get a true comparison of the two models�

With the spline �t model� we are required to use regularization� The spline

�t model can generate a non�smooth model of the O� number density pro�le�

This is undesirable� and regularization smooths the pro�le� Because we use

regularization with this model� an appropriate regularization parameter must

be found� We must therefore do L�curve analysis on this model�

Because of regularization� we cannot generate covariance or correlation matri�

ces from the JTJ matrix of any signi�cance� The variances established by the

covariance matrix would not be the real parameter variance for the establish�

ing of con�dence intervals� We have added a parameter that does not apply

��

��

equally to these intervals� and any covariance matrix is taking this parameter

into account for all its estimates�

We can still use normality tests such as the Anderson�Darling test� �� tests�

and residual regression to check the validity of the �t�

�� The Spline Fit

We use a spline �t for the new inverse procedure� We use a spline

�t of the O� pro�le instead of just using layers because we need �� layers to

generate a reasonable model� By using a spline �t� we generate a parameter set

that is nearly as precise� more easily smoothed� and faster to run� In our spline

�t model� we chose �fteen of the �fty layers� and spline �t the other layers to

end with �fty layers�

To create an accurate spline �t model� we used every other layer near the low

end of the F layer of the ionosphere� about every �� kilometers� starting at

�� kilometers� These close parameters are necessary because the lower end of

the F layer has more change over smaller changes in altitude� and the forward

model is relatively sensitive to the lower altitudes� After the thirteenth param�

eter point� at �� kilometers� we used only every ten layers� or two hundred

kilometers� The upper end of the F layer does not change as quickly and the

forward model is not as sensitive to changes in the upper tail�

At the ends of the spline �ts� it was necessary to add a �nal arbitrary parameter

point to make the tails end where they would end� Without these set spline

points� the upper and lower tails could increase� This increase in O� number

density at the ends of the F layer is not physically reasonable� The lower spline

��

point was set at � kilometers� O� number density of �� The upper spline point

was set at �� kilometers� with an O� number density of ��

Finally� using the spline �t required the transformation of all O� number densi�

ties to their logarithm before �tting the spline� Fitting the spline without �rst

transforming the number densities often resulted in negative number densities

at certain layers as the spline was �t� The cubic polynomial between the two

parameter points required the spline curve to dip into negative numbers when

decreasing rapidly toward zero at the lower altitudes� A logarithmic spline

�t does the same� but the transformation makes negative numbers very small

positives� A negative number density is obviously a physical impossibility� so

this is an undesirable eect� Negative O� number densities result in errors in

the forward model�

�� Forward Model Using the Spline Fit Model

The spline �t model uses a forward model that is nearly identical to

that of the Chapman model� The discrepancy lies at the beginning of the for�

ward model� In the spline �t model� we obviously do not use the Chapman

equation� Instead� we must generate a cubic spline between those points on the

pro�le that are used as parameters in the inverse procedure� Once the spline

�t is generated and �fty layers of the F layer of the ionosphere have been esti�

mated� the forward model proceeds as it did for the Chapman model�

��

�� Spline Inverse Model Using an Initial Ionospheric Model Esti

mated from the Chapman Equation

We used the spline inverse model to model the Chapman equation

produced model� We produced the L curve that showed us what regulariza�

tion parameter worked best� as shown in Figure �� The appropriate inverse

solutions with respect to these regularization parameters are shown in Figures

�� From Figure �� we can see that the appropiate regularization param�

eter to use would be �� At this point we get the bottom left corner of

the L curve� This gives us a best value for �� and the regularization of the �t�

so that both are minimized without increasing the other needlessly�

At a regularization parameter of �� the �� value is �� and the reg�

ularization value is �� The �� value gives us a p�value of �� which

cannot be rejected� The �t passes the �� test� When we check the �t against

the Anderson�Darling normality test� we get a p�value of �� which is also

acceptable� We once again check the ��" con�dence intervals of the parame�

ters of a regression �t of the residuals to see if they are signi�cantly dierent

from zero� The con�dence interval of the constant parameter of the regression

�t is �� This interval contains zero� Thc con�dence interval of the

x coecient is �� which also includes zero� The parameters of the

regression coecients� therefore� also pass the test for a good �t� Finally� we

must look at the standardized residuals of the �t in Figure �� to see if any

recognizable patterns occur�

��

Figure �� L Curve Produced in Generating Fits for the Chapman EquationO� Pro�le

Figure �� Spline Fit of the Chapman Equation Produced O� Number DensityPro�le Using Regularization Parameter ��

��



��

908070605040302010 0

3

2

1

0

-1

-2

C2

splin

e re

sid

R-Sq = 0.3 %

Y = -1.5E-01 + 2.02E-03X

Regression Plot

Figure �� Standardized Residuals of the Spline Fit to the Chapman EquationGenerated Data Set for Regularization Parameter ��

��

Figure �� L Curve for the Spline Model Fit of PIM Pro�le

�� Spline Inverse Model Using an Initial Ionospheric Model Esti

mated from PIM

We also attempted to �t the PIM pro�le with the spline �t inverse

model� The results showed that work could be done on the spline �t model�

The L curve shown in Figure �� shows the poor �ts gotten by this inverse

model� Figures �� show how poorly the �ts estimate the pro�le�

We can evaluate two of the best above models to show that none of these �ts

is any good at all� Figure �� has the best numbers on the L curve� Both the

�� value and regularization value are the lowest of the �ts performed� Looking

at the �gure� it is obvious that the �t is poor� The �gure follows no patterns

that match the behavior of O� ions in the F layer� But even this �t cannot be

good �ts because Figure �� has a �� value of �� This �� value gives an

eective p value for this model of zero� There is no way that these �ts are any

good at all�

��

Figure �� Spline Fit of the PIM Pro�le Using Regularization Parameter��


��


Figure �� Spline Fit of the PIM Pro�le Using Regularization Parameter ��

��



Chapter �

Results and Conclusions

The results of the spline �t model are inconclusive� For the Chap�

man model� the spline �t model appears to work� approximately as well as

the Chapman model� As shown with the PIM model� it does need some work�

With more research the spline �t model could be a solution to the problem of

modelling ionospheric F layer O� number density from �� A airglow� The

conclusive results of this thesis involve the analysis of �ts to data�

After using the Chapman model to form an estimation of the O� number den�

sity in the ionosphere� we must analyze the results using multiple techniques�

The �t could have a very good �� value� but not be a good estimation of the real

F layer� We must use supplementary techniques such as the Anderson�Darling

normality test� residual regression to check for a good residual �t� and exami�

nation of the residuals for strange and unusual patterns suggesting a poor �t�

Using these tests� we can sometimes disclude a �t that has a decent �� value

because of a failure of the �t in some other way�

If we had more accurate sensors� we could have more information in the de�

termination of whether the �t is good or not� After reducing the noise of the

PIM data set by ��"� we showed that nearly twice as many models were poor

�ts� A better� more accurate sensor could reduce noise and make the process

of generating a �t to a data set more reliable�

��

References

�Anderson and Meyer� �� Anderson� D� E� and Meyer� R� R� �� The

OII �� A dayglow� A general model for excitation rate and intensity calcu�

lations� Planetary Space Science� ��!��

�Bain and Engelhardt� �� Bain� L� J� and Engelhardt� M� �� Intero�

duction to Probability and Mathematical Statistics� Duxbury Press� Belmont�

California� second edition�

�Bard� �� Bard� Y� �� Nonlinear Parameter Estimation� Academic

Press� Inc�� New York� New York�

�Bevington and Robinson� �� Bevington� P� R� and Robinson� D� K�

�� Data Reduction and Error Analysis for the Physical Sciences�

WCB#McGraw�Hill� New York� New York� second edition�

�Christiansen et al�� Christiansen� A� B�� Kayser� D� C�� Pranke� J��

Strauss� P� R�� Gutierrez� D� J�� Chakrabarti� S�� McCoy� R� P�� Meier� R� R��

Wolfram� K� D�� and Picone� J� M� �� Instrumentation on the remote

atmospheric and ionospheric detection system experiment� Extreme ultra�

violet spectrometer� photometer� and near�infrared spectrometer� Optical

Engineering� ��

�Daniell Jr� et al�� Daniell Jr�� R� E�� Brown� L� D�� Anderson� D� N��

Fox� M� W�� Doherty� P� H�� Decker� D� T�� Sojka� J� J�� and Schunk� R� W�

��

�� Parameterized ionospheric model� A global ionospheric parameteri�

zation based on �rst principles models� Radio Science� ��!��

�D�Augostino and Stevens� �� D�Augostino� R� B� and Stevens� M� A�� ed�

itors �� Goodness of Fit Techniques� Mercel Dekkar� New York� New

York�

�Gill et al�� Gill� P� E�� Murray� W�� and Wright� M� H� �� Practical

Optimization� Academic Press� Inc�� San Diego� California�

�Hansen� �� Hansen� P� C� �� Regularization Tools� A Matlab Package

for Analysis and Solution of Discrete Ill�Posed Problems� Danish Computing

Center for Research and Education� Building �� Technical University of

Denmark� DK�� Lyngby� Denmark� �� edition� Revised March ��

�Hedin� �� Hedin� A� E� �� MSIS�� thermospheric model� Journal of

Geophysical Research� ��A��!��

�Lawson and Hanson� �� Lawson� C� L� and Hanson� R� J� �� Solving

Least Squares Problems� Prentice�Hall� Englewood Clis� New Jersey�

�Levenberg� �� Levenberg� K� �� A method for the solution of certain

non�linear problems in least squares� Quarterly of Applied Mathematics�

��!��

�Marquardt� �� Marquardt� D� W� �� An algorithm for least�squares

estimation of nonlinear parameters� SIAM Journal of Applied Mathematics�

��!��

��

�McCoy et al�� McCoy� R� P�� D� E� Anderson� J�� and Chakrabarti� S�

�� F� region ion densities from analysis of O� �� A airglow� A para�

metric study with comparisons with satellite data� Journal of Geophysical

Research� ��A��!��

�McCoy et al�� McCoy� R� P�� Dymond� K� F�� Fritz� G� G�� Thonnard�

S� E�� Meier� R� R�� and Regeon� P� A� �� Special sensor ultraviolet

limb imager� An ionospheric and neutral density pro�ler for the Defense

Meteorological Satellite Program satellites� Optical Engineering� ��

�Parker� �� Parker� R� L� �� Geophysical Inverse Theory� Princeton

University Press� Princeton� New Jersey�

�Picone et al�� Picone� J� M�� Meier� R� R�� Kelley� O� A�� Dymond� K� F��

Thomas� R� J�� Melendez�Alvira� D� J�� and McCoy� R� P� �� Investiga�

tion of ionospheric O� remote sensing using the �� A airglow� Journal of

Geophysical Research� ��A��!��

�Ramsey and Schafer� �� Ramsey� F� L� and Schafer� D� W� �� The

Statistical Sleuth� A Course in Methods of Data Analysis� Duxbury Press�

Belmont� California�

�Tikhonov and Arsenin� �� Tikhonov� A� N� and Arsenin� V� Y� �� So�

lutions of Ill�Posed Problems� John Wiley� New York� New York�

�Tikhonov et al�� Tikhonov� A� N�� Goncharsky� A� V�� Stepanov� V� V��

and Yagola� A� G� �� Numerical Methods for the Solution of Ill�Posed

Problems� Kluwer Academic Publishers� Dordrecht� The Netherlands� Orig�

inally published in Moscow� U�S�S�R� in ��

��

�Twomey� �� Twomey� S� �� Introduction to the Mathematics of In�

version in Remote Sensing and Indirect Mesurements� Dover Publications�

Inc�� Mineola� New York�

��

This thesis is accepted on behalf of the faculty of the Institute by the following

committee�

Advisor

Date

Ionospheric Stinger - euler.nmt.edueuler.nmt.edu/~brian/students/jstinger.pdf · Chapman PIM Fit...

Documents

Transcript of Ionospheric Stinger - euler.nmt.edueuler.nmt.edu/~brian/students/jstinger.pdf · Chapman PIM Fit...